Methodology
How the Unstable Index is built, what it measures, and where it falls short.
What the index is
The Unstable Index is a single 0-100 reading of structural fragility in the stablecoin and crypto-collateral system. Higher means more fragile: more leverage, thinner pegs, more crowded positioning, more macro pressure. It is assembled from eight sub-indices, each built from public market and on-chain data.
Leading vs. coincident
The eight sub-indices are split into two groups, and this split is the core of the design:
- Leading signals tend to build before stress becomes visible in prices - leverage accumulating quietly, capital rotating, macro conditions tightening.
- Coincident signals fire during or just after an event - a peg breaking, liquidations cascading, volatility exploding.
They are reported separately as Fragility (leading only) and Stress (coincident only), alongside the combined Overall score. A leading indicator that is still screaming on the day of the crash is not leading, so we hold those components to that standard: several of them are designed to fall as an event turns coincident.
Inputs
| Sub-index | Type | Weight | What it measures | Primary sources | Refresh |
|---|---|---|---|---|---|
| Peg Dispersion | COINCIDENT | 20% | Supply-weighted deviation of tracked stablecoins from $1.00, supply contraction, pool imbalance | DeFiLlama, on-chain pool state, exchange spot | Hourly |
| Leverage & Funding | LEADING | 20% | Perpetual funding rates, open-interest velocity, and a "compressed spring" term: leverage building while volatility stays suppressed | Binance, Bybit, Hyperliquid, Deribit | Hourly |
| Liquidation Proximity | COINCIDENT | 15% | Positioning extremes - how crowded long or short the market is relative to a neutral base | Exchange long/short account ratios | Hourly |
| Volatility | COINCIDENT | 15% | DVOL level, implied-vol term structure (backwardation = acute stress), put/call ratio | Deribit options | Hourly |
| Security | COINCIDENT | 10% | Rolling 30-day USD value lost to exploits, on a logarithmic scale so a single large hack does not pin the reading at the maximum | Public incident feeds | Daily |
| On-chain Flows | LEADING | 8% | Aggregate USD-stablecoin net outflows and supply contraction - redemptions drain the system before prices move | DeFiLlama, chain data | Daily |
| Narrative | COINCIDENT | 7% | News severity composite: exchange halts, exploits, macro drama | Curated RSS feeds | Hourly |
| TradFi Spillover | LEADING | 5% | MOVE index, VIX and VIX/VIX3M backwardation, 20-day dollar rate-of-change, high-yield credit spreads | Yahoo Finance, FRED (St. Louis Fed) | Daily |
The Overall score is the weighted sum above, multiplied by a contagion amplifier that raises it when several sub-indices are stressed at the same time - a crisis in one corner is less dangerous than the same readings everywhere at once.
What stays in-house
The weights, inputs, sources and their direction are all published here - enough to judge whether the construction is sound or to argue with it. What stays in-house is the calibration inside each sub-index: the curves that turn a funding rate or a peg deviation into a 0-100 reading, and the asset-level model behind Peg Dispersion. That is the part that took the work.
Update frequency
- Sub-index inputs are scraped continuously, most on an hourly cycle (see the table above).
- The composite score is computed once per day at 08:00 UTC. That is the number shown on the site and served by the public source feed.
- Day-over-day changes are always measured against the previous day's 08:00 UTC snapshot - never against an intraday value.
- A service restart does not write an extra snapshot. There is exactly one composite reading per day, so the history cannot be inflated by redeploys.
Missing-data policy
This is where most composite indices quietly lie. If a feed goes down and the scorer defaults to zero, the index reads calm precisely when it has gone blind.
- Absent funding, OI or options data → the affected term returns a neutral constant, not 0.
- Absent macro tickers (MOVE, VIX) → a neutral fallback, deliberately above the calm end of the range.
- Sub-indices are weighted over available inputs only, so a dead feed dilutes rather than silently zeroes the score.
- A sub-index with no measurable inputs at all - for example Peg Dispersion when no stablecoin could be scored - is left out of the Overall score for that day and the remaining weights are rescaled, rather than being entered as zero.
Robustness
Historical backtest on real data
In October 2026 the current construction was run day by day over real public data from April 2021 to today: Binance funding, Bybit open interest and long/short ratios, Deribit DVOL, DeFiLlama supply, prices and exploit losses, FRED VIX, dollar index and high-yield spreads. The news signal, the implied-vol term structure, the put/call ratio and the MOVE index have no public archive and sit at their neutral fallbacks, so every reading below is a lower bound on what the live index would have shown.
| Level | Range | Share of days 2021-2026 |
|---|---|---|
| LOW | 0-24 | 79% |
| MODERATE | 25-37 | 17% |
| ELEVATED | 38-49 | 3.3% |
| HIGH | 50-100 | 0.4% |
| Event | T−7 | T−1 | T | T+1 | T+3 | Highest in the 30 days before |
|---|---|---|---|---|---|---|
| May 2021 deleveraging (19 May) | 30 | 32 | 46 | 41 | 35 | 41, ELEVATED from T−8 |
| Terra / UST (9 May 2022) | 21 | 23 | 32 | 53 | 60 | 29 |
| FTX (8 Nov 2022) | 30 | 21 | 17 | 39 | 20 | 49, ELEVATED from T−28 |
| SVB / USDC (10 Mar 2023) | 18 | 16 | 15 | 16 | 22 | 21 |
| Aug 2024 carry unwind (5 Aug) | 18 | 22 | 19 | 21 | 18 | 25 |
| Oct 2025 liquidation day (10 Oct) | 20 | 15 | 14 | 21 | 18 | 21 |
Readings are the 08:00 UTC snapshot, so an event that hits later in the day shows up at T+1. Two honest misses: the SVB weekend stays LOW on on-chain data alone (daily prices understate an intraday depeg, and the replay has no news feed and no Curve pool data, both of which the live index sees), and the October 2025 liquidation day had no leverage build-up signature in these inputs at all. Beyond the named events, the ELEVATED days cluster in the leverage build-ups before the September 2021, December 2021 and March 2024 drawdowns. Across all days, a reading at MODERATE or above was followed by a 10%+ BTC drawdown within a week about twice as often as a LOW reading (26% vs 13%). ELEVATED and HIGH do not separate further on that metric, and BTC implied volatility alone (DVOL) is a stronger drawdown predictor than the composite. The index is a fragility gauge, not a forecast.
Regression suite
Every calibration change runs against a suite before it ships. Since v3.0 the suite re-scores the real daily inputs around the six events above and four calm stretches with the current scorers, and asserts both directions: the index elevates on the events, the leading signals load before them, and the calm stretches stay LOW on median and never reach ELEVATED. Earlier versions used hand-written "event-typical" input ranges; the real-data run showed those were never observed together on any single day, which is why the suite was rebuilt on real inputs. A calibration change that breaks any of these does not get deployed.
Self-backtest
The backtest measures depegs as a discount against each asset's own recent median, not against a flat $1.00 - an asset that trades at $0.998 all year is not depegging every day, and scoring it that way would manufacture flattering hit rates. Over the most recent 99-day window that yields three genuine depeg events. The index caught the systemically relevant one at a maximum score. The two misses were micro-cap assets too small to move the composite.
Known limitations
We would rather state these than have them found for us.
- Micro-cap blind spot. Supply weighting means a small stablecoin can break its peg without meaningfully moving the index. That is intentional - the index measures systemic fragility, not every individual failure - but it does mean it is the wrong tool for monitoring a small position.
- Exogenous shocks are not forecast. The October 2025 liquidation day and the August 2024 carry unwind show no build-up signature in the inputs; the index reads them after the fact, if at all. It is built to see leverage and peg fragility accumulating, not a tariff headline.
- Short live history. The index in its current form has a limited public track record. Historical validation runs over real daily data since 2021, with several inputs at neutral fallbacks, which is not the same as having been live through those events.
- Centralised-venue dependency. Funding, open interest and options data come from a handful of major venues. If those APIs degrade or change shape, coverage narrows.
- No claim of causality. The sub-indices are correlated with stress, not proven causes of it. A rising score is a reason to look closer, not a mechanism.
Changelog
Construction changes are listed here, so a move in the score can be told apart from a change in how it is measured.
- v3.0 - 27 Oct 2026. Calibration anchored on a day-by-day backtest over real data since April 2021 (see Robustness). Three scales had been sized for extremes that never occurred: the funding component now reads 100 at 0.05% per 8 hours (the five-year 99th percentile) instead of 0.1%, which no day reached; peg dispersion uses a supply-weighted root-mean-square so a large coin breaking registers (it is what the API description had always said); TradFi anchors moved from 2008-sized levels to the 2021-2026 range. The contagion amplifier rises to 1.5× when three or more sub-indices are stressed (was 1.25×), since a weighted mean of eight channels dilutes exactly the co-movement that defines systemic stress. The level thresholds moved from 45/65 to 38/50 for ELEVATED/HIGH, set from the backtest distribution: under the old thresholds HIGH was reached on zero real days in five years. Readings from this date are a few points higher than they would have been under v2.2 on the same inputs; the 90-day context line accounts for that within a few weeks. The regression suite now re-scores real event-day inputs instead of hand-written ranges.
- v2.2 - 26 Oct 2026. Four construction fixes after 106 live days, none of which changes the weights. (1) The implied-vol term structure inside Volatility had read "flat" on every snapshot since launch, because the options feed does not carry expiry timestamps; expiry is now parsed from the instrument name and only near-the-money strikes are compared. Calm contango days read a few points lower on Volatility, backwardation days higher. (2) Narrative was the plain mean over about 45 tracked assets, which never left single digits because most assets have no news on a given day; it is now the mean of the three most news-stressed assets. Typical days read roughly 8 points higher on that sub-index, about half a point on the Overall. (3) Inside Leverage & Funding, a fast open-interest unwind now counts half; only build-up is a leading signal, the flush is the cascade itself and is already measured by Volatility and Liquidation Proximity. (4) A sanity bound on the 20-day dollar move in TradFi, after a bad tick once read as a 17 percent move. The API and the live page now also show where today's reading sits in its own last 90 days, because the Overall is deliberately calm-disciplined (live range so far about 10 to 38) and the raw number alone does not say whether a day is unusual.
- v2.1 - 25 Sep 2026. Security moved from a linear to a logarithmic scale. The linear scale saturated after a single large hack and held the sub-index at its maximum for the whole 30-day window. Typical months read the same as before; only the top end changed. The first reading after the change was about 3 points lower on the Overall score than it would have been under the old scale. Security also stopped falling back to zero when its data feed fails.
- 18 Sep 2026. The single-channel warning threshold was lowered from 85 to 78. The Leverage & Funding sub-index cannot reach 85 even at historic extremes, which made the warning unreachable for it.
Questions and corrections
If something here looks wrong, we would rather hear it. Reach us at
@usdus_official.
Methodology version v3.0. This page is updated when the construction changes, not when the score does.